Why is Fingerprint-based indoor localization
still so hard?
Brieuc VIEL, Mikael ASPLUND
Department of Computer and Information Science
Linköping University, Sweden
brivi621@student.liu.se
mikael.asplund@liu.se
Abstract – Wireless indoor localization
systems
and
especially
signal
strength
fingerprinting techniques have been the subject of
significant research efforts in the last decades.
However, most of the proposed solutions require a
costly site-survey to build the radio map which can
be used to match radio signatures with specific
locations. We investigate a novel indoor localization
system that addresses the data collection problem
by progressively and semi-autonomously creating a
radio-map with limited interaction cost. Moreover,
we investigate how spatiotemporal and hardware
properties-based variations can affect the RSSI
values collected and significantly influence the
resulting localization. We show the impact of these
fluctuations on our system and discuss possible
mitigations.
I. INTRODUCTION
The market for location-based services has
increased dramatically in the last few years and is
projected to grow further in the coming years. Today,
mobile devices are often able to determine their
position with an accuracy of just a couple of meters
under favorable conditions. However, in large
buildings the location capabilities drop significantly
due to unavailability of satellite signals (i.e., GPS).
This gap is well known and many attempts have been
made towards achieving indoor localization with
acceptable accuracy, including major players such as
Google, Bing and also Nokia. Despite these efforts, the
indoor localization problem remains largely unsolved.
There are three basic methods that can be used
for localization (in isolation or in combination). Multilateration or multi-angulation (often are both referred
to as triangulation) uses estimated distance or angle to
fixed reference points to calculate the position, scene
analysis
uses
knowledge
about
particular
characteristics (such as RSSI values, or image data) for
a particular location to match with current sensor
values, and finally inertial navigation, which relies on
motion sensors to derive the current location. For
indoor localization, scene-analysis have been deemed a
promising approach since most public and commercial
buildings today have a number of Wi-Fi access points
which can be used to create a sort of signal strength
fingerprint for a given location.
There are many successful implementations of
this method [1, 2, 3, 4, 5] providing a positioning
accuracy of as good as 1.5m. However, scene analysis
and fingerprinting by definition relies on having
collected a sufficient amount of information
beforehand in order to match the current sensor
readings with a previously recorded location. Without
this information, a fingerprinting-based approach will
not be able to provide any location information to the
user. Unfortunately, collecting this information is both
time-consuming and error prone.
In this paper we take a closer look at the basic
conditions
for
indoor
localization
through
fingerprinting. In particular we focus on the collection
and sharing of fingerprints among multiple devices.
We present a novel approach for non-intrusive user
input to match fingerprints with physical locations.
Moreover, we discuss a previously neglected issue
related to the sharing of fingerprints among multiple
devices. Many works assume that the fingerprint made
by one device can be used by another device. Despite
the presence of results showing that this is not the case
[6], this continues to be used as an implicit
assumption. In this paper we strengthen previous
results and show that RSSI variations can occur also
between identical models of popular high-end
smartphones. We show that our system still achieves
acceptable results despite this difference, but we
believe that it is a problem which needs further study
by the research community.
The contributions of this paper are two-fold.
First, we present a novel fingerprinting-based indoor
localization system which provides a reasonable tradeoff between the amount of user interaction and quality
of service. Second, we show that the quality of
information of RSSI signals is less than what is
implicitly assumed by many similar approaches.
The rest of this paper is organized as follows. Section
II describes related works. Then in Section III, we
present the system we designed. Section IV studies the
impact of the RSSI variations. Finally, Section V
concludes this paper.
II. RELATED WORK
Researchers
have
proposed
different
approaches to determine device location in various
indoor environments. Cell-phone manufacturers have
proposed to increase GPS performance by installing
GPS-repeater modules inside buildings to provide
more accuracy to GPS devices. These modules are still
too expensive and do not consider complicated signal’s
propagation inside buildings. Then either large empty
RSSI (dBm)
-55
Signal 1
-75
Signal 2
Signal 3
-95
1
51
101
151
201
251
301
351
Measurements
401
451
501
551
Fig. 1. Wi-Fi RSSI variations over the time at fixed position
room or huge number of repeaters is required to
provide high accuracy[2].
WLAN-based fingerprinting is another
interesting technique proposed in solving the indoor
localization problem. It proposes to use the received
signal strength and properties from different access
points to create a unique combination of data.
Therefore, a constraining preliminary phase is required
to collect fingerprints in the specified area. Such a
system doesn’t manage environment changes like
moving and unpredictable obstacles which may
significantly deteriorate the relevance of the data set.
Several predictive algorithms and probabilistic
models have been implemented to achieve the
matching between measured data and recorded
fingerprints. A wide range of wireless technologies has
also been exploited like Wi-Fi[8] network, GSM[9] or
Bluetooth [10], but also environmental characteristics
such as: color, light and sound [2].
Veljo Otsason [9] proposed an indoor GSMbased fingerprinting localization system with median
accuracy of 5 meters in a large multi-floor building. To
determine the actual location, the measured signals
strengths are compared to the recorded data using the
Weighted K-Nearest Neighbor algorithm to define the
closest estimated position. This GSM-based system
has shown good results and high accuracy in different
types of multi-floor building. RADAR [8] project was
the first attempt to Wi-Fi-based fingerprinting.
Contrary to the other project existing at that time (year
2000), RADAR doesn’t require any special equipment,
it only uses already deployed wireless networks. As
Otsason’s system, a constraining data collection phase
is needed before utilization. Matching is done using
the Nearest Neighbors in Signal Space algorithm.
However, these two systems present significant
limitations: the long training and the low adaptability
to environmental changes (wall modification, new
furniture or addition of a new wireless network). In
this case, a total reconfiguration of the fingerprints
data set is required.
A major innovation was presented by
SurroundSense (2009) [2], which proposed to combine
GSM, Wi-Fi and Bluetooth characteristics with
environmental properties like light, color and sound
intensities to achieve better accuracy. The resulting
fingerprints present a higher uniqueness compare to
other systems but it also complicates the data
collection and matching.
In 2009, PILS [3], a GSM and Wi-Fi
combined project presented an important innovation
while proposing user-friendly system that limit user
interaction and reduce training-phase. The collection
phase is now partially integrated in the running-time.
This is the first step to site-survey suppression and
user-friendly data-collection, providing fast and easy
set-up and configuration. However, time is still
required before the database get enough fingerprints to
perform really efficient localization. Even if the global
method is similar, our implementation is slightly
different, focus is given on data collection and
fingerprint generation to provide more reliable and
stable signals information. So we defined a new and
updated technique record more representative RSSI
values by simplifying this data collection and reducing
the user-interaction needed.
More recently, Fingerprinting method has been
combined with Pedestrian Dead Reckoning (PDR)
technique to provide an hybrid localization
approaches. Pazl [7] presents this innovative
combination, but even with sufficient number of
samples accurate location estimation is still difficult
because the system only use one fingerprint and has to
proceed rapidly to combined with PDR. Moreover, this
system can be quite expensive from an energy
consumption perspective if solely relied on for
continuous location tracking.
Fig. 2. System Architecture
III. SYSTEM DESCRIPTION
In this section, we give an overview of our
system, we explain our motivation, present the main
functionalities and explain the main principle and
algorithms designed. We propose a novel adaptive
indoor localization system dedicated to Android
smartphones based on Radio-Frequency infrastructures
such as GSM, 3G and Wi-Fi networks to provide room
recognition and localization.
Average Euclidean Distance Vs Success Rate
1
0,8
0,6
0,4
0,2
0
Average ED
(dBm)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Localization attempts per Location
Fig. 3. Average Euclidean Distance and Success Rate computed during the evaluation
Xperia
Nexus
Nexus
Galaxy SII
-10
-30
-50
-70
RSSI (dBm)
-10
-30
-50
-70
RSSI (dBm)
RSSI (dBm)
-10
-30
-50
-70
Wi-Fi signals
1 2 3 4 5 6 7 8 9 10
Wi-Fi signals
1 2 3 4 5 6 7 8 9 10
Wi-Fi signals
1 2 3 4 5 6 7 8 9 10
SII #1
SII #2
Fig. 4-6. Wi-Fi RSSI pair analysis
A – Main principle
We designed a system which should work with any
Android smartphone without distinction of brand or
model; for this reason we had to apply some strong
requirements. Based on hardware properties [9] and
analysis [4] we choose to control the number of signals
recorded per fingerprint to a maximum of 6 GSM and
10 Wi-Fi signals per fingerprint. According to related
studies [5], uniqueness of fingerprint and probability
of successful increase significantly with size up to 6
signals. GSM and Wi-Fi both share common properties
like ID and RSSI values. They also present the same
response to external factors and parameters that may
cause significant variation to signal strength value. We
observe that not only does the signal strength of GSM
and Wi-Fi (Figure 1) present huge variations over time
(a well-known fact) but also depending of the
individual device used. Figure 1 illustrates the
variations of three Wi-Fi signals strength (RSSI) at a
fixed position over a period of 10min (1 sample per
second). Unpredictable variations may affect signal
strength, produced by simple environmental changes
and factors, i.e. the number of people present or
moving around the building, the walls, floors, ceilings
or other architectural constraints [11], or simply the
position of the user when taking the measurement.
Even a simple weather condition change, may cause
significant variations [12]. Properties of the device
used may also affect the measurements: variations
have been noticed even between phone of same brand
and model. To reduce the impact of the variation of
received signal strength, we collect several RSSI data
per fingerprint and apply a smoothing algorithm to
limit exposition to strong and sudden variations. This
method results in more representative data, despite the
possible environmental or materiel variations.
The system is organized around four basics
modules shown in Figure 2: the User interface (UI)
module, the Localization module (fingerprint creation
and localization algorithm), the Scanner module
(GSM, Wi-Fi and motion captors) and finally the
Database module. Due to hardware limitations ,
smartphones usually process and record only the six
best GSM neighbor networks [14, 15]. Since our
system is designed for commercial Android
smartphones without any restrictions (except casual
sensors like Wi-Fi, GSM and Accelerometer), we
chose to consider a maximum of six GSM signals per
fingerprint. On the other hand, Wi-Fi interface does
not have such a restriction and may return up to fifty
different networks. Previous work [9, 15] show that the
localization efficiency is linked to the size of the
fingerprint used, based on evaluations done on
different localization systems and methods. The error
rate usually strongly decrease when size increases up
to 10 signals and then decrease slowly until size equals
15 signals. After 16, the error rate becomes rather
stable, implying that no more performance increase is
provided by adding more signals. Based on this study,
we chose to define the size of our fingerprint to 6 GSM
signals and 10 Wi-Fi signals (i.e., a total of 16 signals).
The temptation is high to collect as much signals as
possible for each fingerprint but this would lead to
higher measurements, computation and storage costs
without significantly increase the localization
accuracy.
Fingerprint is related to a relative location.
This position is defined by Cartesian coordinates X
and Y representing the position on a map
corresponding to the current building area. These
coordinates are relative coordinates; they only make
sense on the corresponding map (Cartesian plans) and
would be meaningless if used on another map or even
with a deformed map. We assume that the recognition
of the corresponding map and the possibility to
manage (add and modify) a map in the system does not
affect the accuracy or the precision of the system.
For these reason, map management has been
considered but not implemented. In order to save the
collected data, we defined a database that contains the
data relative to the fingerprints (signal ID + RSSI) and
to the locations (x, y coordinates + related map). The
system uses a local database stored in user’s
smartphones, which contains all the data needed to be
localized.
B – Three-steps algorithm
In order to propose a crowd-sourcing-based
autonomous system, free of site-survey and requiring a
minimized human interaction, we have designed a
novel localization algorithm. The general process is the
same as in other indoor localization process: when a
localization attempt is launched, the system will record
the fingerprint of the current network environment that
will become the reference data. This reference
fingerprint (R-Fingerprint) will be compared to the
data already recorded into the database to find a
possible matching and related location. Our
localization algorithm is divided into three steps in
order to achieve a more representative data collection,
data processing and accurate localization estimation. In
the first step, the system performs a pre-selection
among the previously recorded fingerprints in the
database set, focusing only on relevant data. For each
fingerprint from the database, the system analyzes the
number of signals in common with the reference
fingerprint (RF) (based on signal ID) and keeps only
those respecting a predefined threshold. By this
method, we ensure that every fingerprint the system
will compare has at least a predefined proportion of
signal IDs in common with the R-Fingerprint. To
ensure the accuracy of our comparison, the preselected fingerprints will be temporary truncated to
keep only the common signals. So every pre-selected
data has the same size and the same percentage of
corresponding signals with the RF.
The second step is the comparison between
resulting set of step one and reference fingerprint. The
comparison is done based on the Euclidean Distance
(ED) calculated between the strength value of each RF
signals and the other fingerprint signals. The unit of
ED is dBm per signal. If no satisfying location is
returned, the system proceeds to step three.
The Weighted k-Nearest Neighbor estimation
method is applied in step three to evaluate the current
location based on the x and y position of the k closest
fingerprints based on their respective ED. A weight is
also applied to the position of each of the k-nearest
neighbors based on its ED.
The innovation of this algorithm lies first in
the way it selects the data to process, making the result
more accurate by comparing RSSI data of same
common size fingerprints. We insure that the
fingerprints compared have exactly the same number
of signals. Novelty is also presented by the algorithm
adaptation capacity and the flexibility of the result
returned. Adaptation and flexibility are key points in
indoor localization, since environmental changes are
most likely to happen and impact the wireless signal
propagation and so the corresponding fingerprint. Even
the morphology of the user or the way he carries his
phone may have an impact. In this case and in
comparison with existing systems, our system will
automatically adapt itself and overpass this change by
progressively updating its database. The flexibility of
our algorithm makes it possible since it doesn’t return
a perfect value but a matching state based on the ED
calculated. These three states are the Perfect Matching
(we assume the returned location is true), the Possible
state (confirmation is asked) and the Estimation State
(step three of the algorithm). With these states, the
system is able to manage environmental changes and
limit the impact on the resulting localization.
C – Data Collection and interval scanning
We propose to suppress the site-survey phase
and include data collection into the normal utilization
process. The interactions required are then limited and
no offline phase is required. Data collection is
accomplished by two different methods.
First, by giving the opportunity to add a new location
in the database or correct the proposition of the
localization process.
The second way to collect data is the interval
scanning module. Based on acceleration magnitude
measurements, the system applies a motion detection
(Fig. 7). When the device (and by extension the user)
stays stationary over a period of time, the system
performs a continuous data collection. If the
fingerprint of the current position is not already
recorded in the database, a new fingerprint is created
compiling the average values collected until the user
move away. Then the system asks the user to relate the
fingerprint with his previous location on a map. The
basic idea behind this behavior is that it is more
convenient for user to remember his previous location
during a period of time he stayed stationary rather than
providing it precisely at a certain time.
Fig. 7. Motion detection and interval scanning
This system provides more fingerprints and
locations to the database with a limited interaction.
This also gives the system the freedom to report the
confirmation to a future moment more convenient for
the user (e.g., moment where he comes back to his
desk/table). A similar system has been presented in
PILS project [3], with less autonomy allowed to the
system and more interactions imposed to the user.
More details are asked to the user about the position to
link the fingerprint to the corresponding location
(name, location, function). Moreover, the PILS system
doesn’t compare the data with the already recorded
ones, creating this way doubles locations and
fingerprints into the database.
Fig. 8. Evaluation area and targeted
locations , IDA building LiU.
D – Evaluation and results
To evaluate the reliability of our system, we
defined a protocol to analyze the result of the
localization process at 10 predefined positions,
performing thirty attempts for each location, so a total
of 300 measurements. This evaluation was done in an
academic building of Linköping University (Fig. 8 800m²), in normal use conditions with full-capacity
and performance. No consideration is given to the size
of the user, to the position the user carries the phone in
since the variations resulting from these parameters
must be override by the system. To analyze the results
of this evaluation, we recorded the closest Euclidean
Distance generated during each localization process,
the database size and the number of user interactions
during the evaluation (Fig. 3). From the results of this
evaluation, we observe two temporary period before
achieving full-capacity. During the first ten
localization attempts, the system collects the first data
to provide localization and progressively fills the
database. Then system keeps collecting data to provide
complete overview of every location. Over the 300
values calculated during this evaluation, we achieved a
success rate of 82% correct localization with roomlevel accuracy and an Euclidean Distance of 0.47
dBm.
However, even after thirty attempts,
localization failure may still happened due to access
point signal punctual disappearance or exceptional
meteorological conditions. By relating results values
and figure 8 together, we can assume that our system is
less efficient in high room-density area (locations 1, 4,
5 and 6) and especially with small size rooms
(locations 5 and 6).
We also observed an unpredicted failure,
“false positive” localization may happen, especially
during fifteen first localization attempts when data
signal map is not complete enough to manage every
situation and in high density areas previously
described. Even if these “false positive” results have
an short-term impact on the localization (false location
is returned to the user), it will be automatically
overwritten by the fingerprints collected at mid-term.
The system will continuously record new data that will
reduce the accuracy and ED of the “false positive”
fingerprint. After evaluation, we define that ”False
Positive” values only represent 4% of the final amount
of localization attempts.
Finally, we investigated the number of user
interactions required during the utilization of our
system. This number of interaction is closely related to
the matching case encountered (1 to 3 interactions
maximum). For each case, we assume that the
launching of a localization attempt (i.e. press the start
button) is considered as one interaction, so in the best
case, this is the only interaction required.
During the evaluation of the system, the
number of interactions applied by the user has been
recorded. The user interacts an average of 1.57 time
with the system per localization attempt. As précised,
in the worst case, the user would have to interact a
maximum of 3 times per attempt. During the last 10
localization attempts corresponding to a normal
utilization context, the average number of interaction
required by the system to the user decreases under 1.3
time per localization.
IV – IMPACT OF RSSI VARIATIONS
The signal strength measured by a given
device may fluctuate significantly depending of
spatiotemporal and hardware-based properties. We
now proceed to describe the impact these variations
have on our system and especially on the matching
resulting from our algorithm.
This experiment consists of measuring
simultaneously GSM and Wi-Fi fingerprints on
different pairs of devices. We compare devices of
different brands (Fig. 4.) devices of same brand but
different models (Fig. 5.) and devices of same model
(Fig. 6.). Four different devices have been considered,
one Samsung galaxy Nexus, one Sony Ericsson Xperia
and two Samsung Galaxy SII; all of them running on
Android OS (v.4.2, API lvl. 17). For each pair of
device, we measured the respective RSSI average per
signal (Fig. 4-6.).
We observe notable differences between data
recorded, even with devices of the same brand and
model. Difference of as much as 12 dBm can be
observed between data from same signal recorded by
different devices. Contrarily to what we expected,
devices from the same brand with chipset from the
same vendor will not record the same value. Figure 6
shows that variation of almost 9dBm may occurs
between these two Samsung mobiles. Even if the size
of our experimental set is not consequent, this further
strengthens the work presented by Lui et al. from
University of New South Wales, Sydney, AU [6].
They showed that different Wi-Fi devices (mostly
wireless network adapters) show a considerable
variation of RSSI readings (including one case of three
identical network adapters). We demonstrate that the
same problem can be seen for identical mobile phones
of a popular model.
Hardware properties-based fluctuations have a
direct impact on RSSI value measured by a device and
by consequence on the Euclidean Distance calculated
during localization process to evaluate fingerprints
matching.
Pair of
devices
Nexus vs.
Xperia
Nexus vs. SII
SII vs. SII
ED (dBm)
0.6833
0.7858
0.3316
Table. 1. Simulated Euclidean Distance between the respective
pair of devices
Table 1 presents the simulated Euclidean
Distance calculated from data collected during pair
devices experiment. Based on evaluations, we assume
that an ED of less than 0.5 dBm corresponds to a
perfect matching, i.e. the location returned has a high
probability to be correct. Between 0.5 and 1.0 dBm,
probability is still positive but confirmation should be
given by the user. So for two fingerprints taken
simultaneously by two different devices in the same
conditions, accuracy of the ED calculated may differ
significantly and the resulting location returned by our
system strongly diverge despite all the precautions
taken for data collection. However, this observation is
moderated when both devices are of same brand and
model.
It seems as if some of the RSSI-differences
between different devices are systematic, which would
suggest that they could be measured, modelled, and
accounted for by the localization algorithms. However,
considering for example Figure 6, we see that the
values recorded by device 1 is higher for some of the
signals and lower for the others compared to the
readings by device 2. This suggest that one would need
a fairly sophisticated learning algorithm to capture the
change in RSSI values caused by an individual device.
However, Figure 4 and 5, where the differences are
similar across all signals, suggests that significant
improvements could likely be achieved even without a
perfect algorithm.
V – CONCLUSIONS
Indoor localization remains elusive. Despite
the wealth of research papers claiming to have found
the solution, something is holding back successful
deployment. We believe that one of the main issues
lies in the ability to collect relevant context
information of high quality about each location. In this
paper we present a prototype system where this type of
data collection is integrated in the localization tool. We
show that this prototype shows acceptable results with
relatively simple algorithms. It would be reasonable to
assume that pairing our approach with more
sophisticated sensor-fusion algorithms could yield
even better results.
However, our evaluations also point out an
inherent weakness with RSSI-based fingerprinting. If
the quality of this information source is so low that
differences of 12 dBm between the averaged data from
two
different devices can be observed, then it into
question how well crowd-sourcing approaches can be
made to work. Obviously, algorithms will need to be
both robust and adaptive to cope with this problem.
More research is needed to more precisely model
variations between devices to be able to account for
this source of inaccuracy.
It should be noted that all the data used in the
work have been collected in a specific academic
building with its particular characteristics and
properties. Data have been measured using a limited
number of devices, with specific environmental
conditions. More experiments will be required to fully
validate the effectiveness of our algorithm. Moreover,
we believe that our observations regarding differences
in RSSI should be extended with more experimental
data.
REFERENCES
[1] B. VIEL, “Adaptive Indoor Localization System for Android
Smartphones”. Linkoping University, 2013.
[2] M. Azizyan, R. R. Choudhury, “SurroundSense: mobile phone
localization using ambient sound and light”. Mobile Computing
and Communications Review 13, 2009.
[3] P. Bolliger, K. Partridge, M. Chu , M. Langheinrich,
“Improving Location Fingerprinting through Motion Detection and
Asynchronous Interval Labeling”, Proceedings of the 4th
International Symposium on Location and Context Awareness,
May 2009, Tokyo, Japan
[4] K. Kaemarungsi and P. Krishnarmurthy, "Modeling of indoor
positioning system based on location fingerprinting", Twenty-third
Annual Joint Conference of the IEEE Computer and
Communications Societies (INFOCOM'04), China 2004.
[5] V. Honkavirta , T. Perala , S. Ali-Loytty and R. Piche "A
comparative survey of WLAN location fingerprinting
methods", Positioning, Navigation Commun., 2009.
[6] G. Lui, T. Gallagher, B. Li, A. G. Dempster, and C. Rizos,
"Differences in RSSI readings made by different Wi-Fi chipsets: A
limitation of WLAN localization," international Conference, 2011.
[7] V. Radu, L. Kriara, and M. K. Marina. Pazl: “A Mobile
Crowdsensing based Indoor WiFi Monitoring System”. In Proc.
IEEE CNSM, 2013.
[8]] P.Babl, et al, “RADAR: An in-Building RF-based User
Location and Tracking System.” In Proceedings of the IEEE
Infocom, Tel-Aviv, Israel. Mar 2000
[9] V. Otsason, A. Varshavsky, A. LaMarca and E. de Lara,
"Accurate GSM indoor location", Springer-Verlag Berlin
Heidelberg, UbiComp 2005..
[10] A. Bekkelien, “Bluetooth Indoor Positioning”, Master thesis,
University of Geneva, Mar 2012
[11] P. Prasithsangaree, P. Krishnamurthy, P. Chrysanthis, “On
indoor position location with wireless LANs”, in The 13th IEEE
International Symposium on Personal, Indoor and Mobile Radio
Communications, vol. 2, 2002.
[12] C. A. Boano , J. Brown , Z. He , U. Roedig and T.
Voigt "Low-power radio communication in industrial outdoor
deployments: The impact of weather conditions and ATEXcompliance", Proc. 1st Int. Conf. Sens. Netw. Appl., Exp. Logistics
(Sensappeal), 2009
[13] A. Bekkelien, “Bluetooth Indoor Positioning”, Master Thesis,
2012.
[14] V. Otsason, “Accurate Indoor localization Using Wide GSM
fingerprinting”, Master thesis, Tartu University, published 2005.
[15] V. Honkavirta , T. Perala , S. Ali-Loytty and R. Piche, “A
comparative survey of WLAN location fingerprinting
methods”, Positioning, Navigation Commun., published 2009.